Native Bedrock Codex missing explicit prompt cache controls causes high write spend
Developers using native OpenAI Codex CLI via Amazon Bedrock Mantle on GPT-5.6 Sol report high costs driven by missing explicit prompt cache controls. Telemetry shows cache writes accounting for nearly 85% of total model spend due to missing request options.
Impact: High
Why it matters
Inspect your AWS Bedrock CloudWatch telemetry to ensure agentic coding sessions are not re-writing full prompt contexts on every turn.
TL;DR
- 01Native Bedrock Codex currently lacks explicit prompt cache controls for GPT-5.6 Sol.
- 02Cache writes can account for up to 85% of total LLM API spend in agentic sessions.
- 03Monitor per-turn telemetry for zero cached_input_tokens to catch prefix rewrite loops.
Key facts
- Cache Write Spend Share
- ~85% of model spend
- Sample Session Tokens
- 6.709M write tokens / 0 cached tokens
- Average Write Overhead
- ~88K tokens per request
Explicit Prompt Caching Gap on AWS Bedrock
When executing agentic coding workflows using native Codex CLI over the aws-bedrock provider, requests sent to the Bedrock Mantle Responses API fail to opt into explicit prompt caching for gpt-5.6-sol. Consequently, long instruction and tool definitions are continuously written as new cache entries rather than being read from existing prefixes.
Usage Telemetry and Cost Impact
- Production data from August 5 through August 8 showed cache writes accounted for 85% of estimated model spend.
- A benchmark session logged 76 Sol requests consuming 6.709M cache_write_input_tokens and 0 cached_input_tokens.
- Average cache-write overhead measured approximately 88K tokens per request without any client-side CloudWatch errors.
Required Config Remediation
Because Bedrock provider settings in config.toml only expose transport and authentication flags, developers cannot manually inject request-body transformations. Resolution requires adding native support for prompt_cache_options serialization and explicit prompt_cache_breakpoint placement in Codex.
Try it in 2 minutes
# Current config.toml lacks explicit request body prompt_cache_options
[model_providers.aws-bedrock]
region = "us-east-1"
model = "gpt-5.6-sol"toml
✓ When to use
- When orchestrating long agentic coding sessions using Codex CLI over AWS Bedrock provider infrastructure.
- When auditing unexpected AWS Bedrock LLM billing spikes on agentic workloads.
✕ When NOT to use
- When running Codex via direct OpenAI APIs where prompt caching triggers automatically.
- On short single-turn completions without large prompt prefixes or context windows.
What to do today
- Check CloudWatch metrics for high cache_write_input_tokens vs zero cached_input_tokens on Bedrock Codex sessions.
- Track issue #37674 on the official OpenAI Codex repository for explicit prompt_cache_options implementation updates.
What the community says
“"causing" -> "costing", right?”
“I feel like this is the kind of substantial change to your product that you would need to tell your customers about. It would be simply disrespectful to your customers to not disclose this upfront.”
Sources